基于融合边缘变化信息全卷积神经网络的遥感图像变化检测

Remote Sensing Image Change Detection Based on Fully Convolutional Neural Networks with Edge Change Information

高分辨率遥感图像变化检测是了解地表变化的关键,是遥感图像处理领域的一个重要分支。现有很多基于深度学习的变化检测方法,取得了良好的效果,但是不易获得高分辨率遥感图像中的结构细节且检测精度有待提高。因此,该文提出融合了边缘变化信息和通道注意力模块的网络框架(EANet),分为边缘结构变化信息检测、深度特征提取和变化区域判别3个模块。首先,为了得到双时相图像的边缘变化信息,对其进行边缘检测得到边缘图,并将边缘图相减得到边缘差异图;其次,考虑到高分辨率遥感图像精细的图像细节和复杂的纹理特征,为了充分提取单个图像的深度特征,构建基于VGG-16网络的3支路模型,分别提取双时相图像和边缘差异图的深度特征;最后,为了提高检测精度,提出将通道注意力机制嵌入到模型中,以关注信息量大的通道特征来更好地进行变化区域的判别。实验结果表明,无论从视觉解释或精度衡量上看,提出算法与目前已有的一些方法相比,具有一定的优越性。

Change detection in high resolution remote sensing images is the key to understanding of land surface changes.Change detection of remote sensing images is an important branch of remote sensing image processing.Many existing change detection methods based on deep learning have achieved good results,but it is not easy to obtain the structural details of high resolution remote sensing images,and the accuracy of the detection needs to be improved.Therefore,a network framework which combines Edge change information and channel Attention Network module(EANet)is proposed.EANet is divided into three modules:Edge structure change information detection,depth feature extraction and change area discrimination.Firstly,in order to get the edge change information of the two-phase images,the edge of the two-phase images is detected to get the edge images,and the edge images is subtracted to get the edge difference images.Secondly,in consideration of the fine image details and complex texture features of high resolution remote sensing images,in order to extract fully the depth features of a single image,a model with three branches based on VGG-16 network is constructed to extract the depth features of bitemporal images and edge difference images respectively.Finally,in order to improve the accuracy of the detection,the channel attention mechanism is embedded into the model to focus on the channel features with large amount of information to identify better the changed regions.The experimental results show that the proposed algorithm is superior to some existing methods in terms of visual interpretation and accuracy measurement.